X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=5a3ca7444946751f4e208a32affdad5fb98051b9;hp=bac65739a1ad05843561e2f7dfca1e389265b769;hb=0ce8b119b65849e537cab628a176a0ec4238aab0;hpb=40e1a6f09be0d487ce08e68e2bf9b0a01d771c8a

diff --git a/io_u.c b/io_u.c
index bac65739..5a3ca744 100644
--- a/io_u.c
+++ b/io_u.c
@@ -8,6 +8,8 @@
 #include "fio.h"
 #include "hash.h"
 #include "verify.h"
+#include "trim.h"
+#include "lib/rand.h"
 
 struct io_completion_data {
 	int nr;				/* input */
@@ -28,7 +30,7 @@ static int random_map_free(struct fio_file *f, const unsigned long long block)
 
 	dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
 
-	return (f->file_map[idx] & (1 << bit)) == 0;
+	return (f->file_map[idx] & (1UL << bit)) == 0;
 }
 
 /*
@@ -40,24 +42,27 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 	struct fio_file *f = io_u->file;
 	unsigned long long block;
 	unsigned int blocks, nr_blocks;
+	int busy_check;
 
 	block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
 	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
 	blocks = 0;
+	busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
 
 	while (nr_blocks) {
-		unsigned int this_blocks, mask;
 		unsigned int idx, bit;
+		unsigned long mask, this_blocks;
 
 		/*
 		 * If we have a mixed random workload, we may
 		 * encounter blocks we already did IO to.
 		 */
-		if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
-			if (!blocks)
-				blocks = 1;
+		if (!busy_check) {
+			blocks = nr_blocks;
 			break;
 		}
+		if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
+			break;
 
 		idx = RAND_MAP_IDX(f, block);
 		bit = RAND_MAP_BIT(f, block);
@@ -68,10 +73,20 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 		if (this_blocks + bit > BLOCKS_PER_MAP)
 			this_blocks = BLOCKS_PER_MAP - bit;
 
-		if (this_blocks == BLOCKS_PER_MAP)
-			mask = -1U;
-		else
-			mask = ((1U << this_blocks) - 1) << bit;
+		do {
+			if (this_blocks == BLOCKS_PER_MAP)
+				mask = -1UL;
+			else
+				mask = ((1UL << this_blocks) - 1) << bit;
+	
+			if (!(f->file_map[idx] & mask))
+				break;
+
+			this_blocks--;
+		} while (this_blocks);
+
+		if (!this_blocks)
+			break;
 
 		f->file_map[idx] |= mask;
 		nr_blocks -= this_blocks;
@@ -89,6 +104,8 @@ static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
 	unsigned long long max_blocks;
 	unsigned long long max_size;
 
+	assert(ddir_rw(ddir));
+
 	/*
 	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
 	 */
@@ -109,22 +126,27 @@ static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
 			       enum fio_ddir ddir, unsigned long long *b)
 {
-	unsigned long long min_bs = td->o.rw_min_bs;
+	unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
 	int i;
 
+	lastb = last_block(td, f, ddir);
+	if (!lastb)
+		return 1;
+
 	i = f->last_free_lookup;
-	*b = (i * BLOCKS_PER_MAP);
-	while ((*b) * min_bs < f->real_file_size &&
-		(*b) * min_bs < f->io_size) {
-		if (f->file_map[i] != (unsigned int) -1) {
-			*b += ffz(f->file_map[i]);
-			if (*b > last_block(td, f, ddir))
+	block = i * BLOCKS_PER_MAP;
+	while (block * min_bs < f->real_file_size &&
+		block * min_bs < f->io_size) {
+		if (f->file_map[i] != -1UL) {
+			block += ffz(f->file_map[i]);
+			if (block > lastb)
 				break;
 			f->last_free_lookup = i;
+			*b = block;
 			return 0;
 		}
 
-		*b += BLOCKS_PER_MAP;
+		block += BLOCKS_PER_MAP;
 		i++;
 	}
 
@@ -135,31 +157,40 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
 				enum fio_ddir ddir, unsigned long long *b)
 {
-	unsigned long long r;
+	unsigned long long r, lastb;
 	int loops = 5;
 
+	lastb = last_block(td, f, ddir);
+	if (!lastb)
+		return 1;
+
+	if (f->failed_rands >= 200)
+		goto ffz;
+
 	do {
 		r = os_random_long(&td->random_state);
 		dprint(FD_RANDOM, "off rand %llu\n", r);
-		*b = (last_block(td, f, ddir) - 1)
-			* (r / ((unsigned long long) OS_RAND_MAX + 1.0));
+		*b = (lastb - 1) * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
 
 		/*
 		 * if we are not maintaining a random map, we are done.
 		 */
 		if (!file_randommap(td, f))
-			return 0;
+			goto ret_good;
 
 		/*
 		 * calculate map offset and check if it's free
 		 */
 		if (random_map_free(f, *b))
-			return 0;
+			goto ret_good;
 
 		dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
 									*b);
 	} while (--loops);
 
+	if (!f->failed_rands++)
+		f->last_free_lookup = 0;
+
 	/*
 	 * we get here, if we didn't suceed in looking up a block. generate
 	 * a random start offset into the filemap, and find the first free
@@ -170,7 +201,7 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
 		f->last_free_lookup = (f->num_maps - 1) *
 					(r / (OS_RAND_MAX + 1.0));
 		if (!get_next_free_block(td, f, ddir, b))
-			return 0;
+			goto ret;
 
 		r = os_random_long(&td->random_state);
 	} while (--loops);
@@ -179,7 +210,76 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
 	 * that didn't work either, try exhaustive search from the start
 	 */
 	f->last_free_lookup = 0;
+ffz:
+	if (!get_next_free_block(td, f, ddir, b))
+		return 0;
+	f->last_free_lookup = 0;
 	return get_next_free_block(td, f, ddir, b);
+ret_good:
+	f->failed_rands = 0;
+ret:
+	return 0;
+}
+
+static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
+			       enum fio_ddir ddir, unsigned long long *b)
+{
+	if (get_next_rand_offset(td, f, ddir, b)) {
+		dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
+				f->file_name, f->last_pos, f->real_file_size);
+		return 1;
+	}
+
+	return 0;
+}
+
+static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
+			      enum fio_ddir ddir, unsigned long long *b)
+{
+	assert(ddir_rw(ddir));
+
+	if (f->last_pos < f->real_file_size) {
+		*b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
+		return 0;
+	}
+
+	return 1;
+}
+
+static int get_next_block(struct thread_data *td, struct io_u *io_u,
+			  enum fio_ddir ddir, int rw_seq, unsigned long long *b)
+{
+	struct fio_file *f = io_u->file;
+	int ret;
+
+	assert(ddir_rw(ddir));
+
+	if (rw_seq) {
+		if (td_random(td))
+			ret = get_next_rand_block(td, f, ddir, b);
+		else
+			ret = get_next_seq_block(td, f, ddir, b);
+	} else {
+		io_u->flags |= IO_U_F_BUSY_OK;
+
+		if (td->o.rw_seq == RW_SEQ_SEQ) {
+			ret = get_next_seq_block(td, f, ddir, b);
+			if (ret)
+				ret = get_next_rand_block(td, f, ddir, b);
+		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
+			if (f->last_start != -1ULL)
+				*b = (f->last_start - f->file_offset)
+					/ td->o.min_bs[ddir];
+			else
+				*b = 0;
+			ret = 0;
+		} else {
+			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
+			ret = 1;
+		}
+	}
+	
+	return ret;
 }
 
 /*
@@ -187,33 +287,23 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
  * until we find a free one. For sequential io, just return the end of
  * the last io issued.
  */
-static int get_next_offset(struct thread_data *td, struct io_u *io_u)
+static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
 {
 	struct fio_file *f = io_u->file;
 	unsigned long long b;
 	enum fio_ddir ddir = io_u->ddir;
+	int rw_seq_hit = 0;
 
-	if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
-		td->ddir_nr = td->o.ddir_nr;
+	assert(ddir_rw(ddir));
 
-		if (get_next_rand_offset(td, f, ddir, &b)) {
-			dprint(FD_IO, "%s: getting rand offset failed\n",
-				f->file_name);
-			return 1;
-		}
-	} else {
-		if (f->last_pos >= f->real_file_size) {
-			if (!td_random(td) ||
-			     get_next_rand_offset(td, f, ddir, &b)) {
-				dprint(FD_IO, "%s: pos %llu > size %llu\n",
-						f->file_name, f->last_pos,
-						f->real_file_size);
-				return 1;
-			}
-		} else
-			b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
+	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
+		rw_seq_hit = 1;
+		td->ddir_seq_nr = td->o.ddir_seq_nr;
 	}
 
+	if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
+		return 1;
+
 	io_u->offset = b * td->o.ba[ddir];
 	if (io_u->offset >= f->io_size) {
 		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
@@ -231,18 +321,25 @@ static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 	return 0;
 }
 
-static inline int is_power_of_2(unsigned int val)
+static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 {
-	return (val != 0 && ((val & (val - 1)) == 0));
+	struct prof_io_ops *ops = &td->prof_io_ops;
+
+	if (ops->fill_io_u_off)
+		return ops->fill_io_u_off(td, io_u);
+
+	return __get_next_offset(td, io_u);
 }
 
-static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
+static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 {
 	const int ddir = io_u->ddir;
 	unsigned int uninitialized_var(buflen);
 	unsigned int minbs, maxbs;
 	long r;
 
+	assert(ddir_rw(ddir));
+
 	minbs = td->o.min_bs[ddir];
 	maxbs = td->o.max_bs[ddir];
 
@@ -281,6 +378,16 @@ static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
 	return buflen;
 }
 
+static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
+{
+	struct prof_io_ops *ops = &td->prof_io_ops;
+
+	if (ops->fill_io_u_size)
+		return ops->fill_io_u_size(td, io_u);
+
+	return __get_next_buflen(td, io_u);
+}
+
 static void set_rwmix_bytes(struct thread_data *td)
 {
 	unsigned int diff;
@@ -313,6 +420,8 @@ static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
 	struct timeval t;
 	long usec;
 
+	assert(ddir_rw(ddir));
+
 	if (td->rate_pending_usleep[ddir] <= 0)
 		return ddir;
 
@@ -350,7 +459,7 @@ static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
 	odir = ddir ^ 1;
 	if (td_rw(td) && __should_check_rate(td, odir))
 		td->rate_pending_usleep[odir] -= usec;
-	
+
 	return ddir;
 }
 
@@ -363,6 +472,30 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 {
 	enum fio_ddir ddir;
 
+	/*
+	 * see if it's time to fsync
+	 */
+	if (td->o.fsync_blocks &&
+	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_SYNC;
+
+	/*
+	 * see if it's time to fdatasync
+	 */
+	if (td->o.fdatasync_blocks &&
+	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_DATASYNC;
+
+	/*
+	 * see if it's time to sync_file_range
+	 */
+	if (td->sync_file_range_nr &&
+	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_SYNC_FILE_RANGE;
+
 	if (td_rw(td)) {
 		/*
 		 * Check if it's time to seed a new data direction.
@@ -390,7 +523,18 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 	return td->rwmix_ddir;
 }
 
-static void put_file_log(struct thread_data *td, struct fio_file *f)
+static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
+{
+	io_u->ddir = get_rw_ddir(td);
+
+	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
+	    td->o.barrier_blocks &&
+	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
+	     td->io_issues[DDIR_WRITE])
+		io_u->flags |= IO_U_F_BARRIER;
+}
+
+void put_file_log(struct thread_data *td, struct fio_file *f)
 {
 	int ret = put_file(td, f);
 
@@ -400,16 +544,27 @@ static void put_file_log(struct thread_data *td, struct fio_file *f)
 
 void put_io_u(struct thread_data *td, struct io_u *io_u)
 {
-	assert((io_u->flags & IO_U_F_FREE) == 0);
+	td_io_u_lock(td);
+
 	io_u->flags |= IO_U_F_FREE;
+	io_u->flags &= ~IO_U_F_FREE_DEF;
 
 	if (io_u->file)
 		put_file_log(td, io_u->file);
 
 	io_u->file = NULL;
-	flist_del(&io_u->list);
+	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
+		td->cur_depth--;
+	flist_del_init(&io_u->list);
 	flist_add(&io_u->list, &td->io_u_freelist);
-	td->cur_depth--;
+	td_io_u_unlock(td);
+	td_io_u_free_notify(td);
+}
+
+void clear_io_u(struct thread_data *td, struct io_u *io_u)
+{
+	io_u->flags &= ~IO_U_F_FLIGHT;
+	put_io_u(td, io_u);
 }
 
 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
@@ -418,15 +573,18 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 
 	dprint(FD_IO, "requeue %p\n", __io_u);
 
+	td_io_u_lock(td);
+
 	__io_u->flags |= IO_U_F_FREE;
-	if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
+	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
 		td->io_issues[__io_u->ddir]--;
 
 	__io_u->flags &= ~IO_U_F_FLIGHT;
-
+	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
+		td->cur_depth--;
 	flist_del(&__io_u->list);
 	flist_add_tail(&__io_u->list, &td->io_u_requeues);
-	td->cur_depth--;
+	td_io_u_unlock(td);
 	*io_u = NULL;
 }
 
@@ -435,17 +593,13 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	if (td->io_ops->flags & FIO_NOIO)
 		goto out;
 
+	set_rw_ddir(td, io_u);
+
 	/*
-	 * see if it's time to sync
+	 * fsync() or fdatasync() or trim etc, we are done
 	 */
-	if (td->o.fsync_blocks &&
-	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
-	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
-		io_u->ddir = DDIR_SYNC;
+	if (!ddir_rw(io_u->ddir))
 		goto out;
-	}
-
-	io_u->ddir = get_rw_ddir(td);
 
 	/*
 	 * See if it's time to switch to a new zone
@@ -496,31 +650,31 @@ out:
 
 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
 {
-	int index = 0;
+	int idx = 0;
 
 	switch (nr) {
 	default:
-		index = 6;
+		idx = 6;
 		break;
 	case 33 ... 64:
-		index = 5;
+		idx = 5;
 		break;
 	case 17 ... 32:
-		index = 4;
+		idx = 4;
 		break;
 	case 9 ... 16:
-		index = 3;
+		idx = 3;
 		break;
 	case 5 ... 8:
-		index = 2;
+		idx = 2;
 		break;
 	case 1 ... 4:
-		index = 1;
+		idx = 1;
 	case 0:
 		break;
 	}
 
-	map[index]++;
+	map[idx]++;
 }
 
 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
@@ -537,117 +691,117 @@ void io_u_mark_complete(struct thread_data *td, unsigned int nr)
 
 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
 {
-	int index = 0;
+	int idx = 0;
 
 	switch (td->cur_depth) {
 	default:
-		index = 6;
+		idx = 6;
 		break;
 	case 32 ... 63:
-		index = 5;
+		idx = 5;
 		break;
 	case 16 ... 31:
-		index = 4;
+		idx = 4;
 		break;
 	case 8 ... 15:
-		index = 3;
+		idx = 3;
 		break;
 	case 4 ... 7:
-		index = 2;
+		idx = 2;
 		break;
 	case 2 ... 3:
-		index = 1;
+		idx = 1;
 	case 1:
 		break;
 	}
 
-	td->ts.io_u_map[index] += nr;
+	td->ts.io_u_map[idx] += nr;
 }
 
 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
 {
-	int index = 0;
+	int idx = 0;
 
 	assert(usec < 1000);
 
 	switch (usec) {
 	case 750 ... 999:
-		index = 9;
+		idx = 9;
 		break;
 	case 500 ... 749:
-		index = 8;
+		idx = 8;
 		break;
 	case 250 ... 499:
-		index = 7;
+		idx = 7;
 		break;
 	case 100 ... 249:
-		index = 6;
+		idx = 6;
 		break;
 	case 50 ... 99:
-		index = 5;
+		idx = 5;
 		break;
 	case 20 ... 49:
-		index = 4;
+		idx = 4;
 		break;
 	case 10 ... 19:
-		index = 3;
+		idx = 3;
 		break;
 	case 4 ... 9:
-		index = 2;
+		idx = 2;
 		break;
 	case 2 ... 3:
-		index = 1;
+		idx = 1;
 	case 0 ... 1:
 		break;
 	}
 
-	assert(index < FIO_IO_U_LAT_U_NR);
-	td->ts.io_u_lat_u[index]++;
+	assert(idx < FIO_IO_U_LAT_U_NR);
+	td->ts.io_u_lat_u[idx]++;
 }
 
 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
 {
-	int index = 0;
+	int idx = 0;
 
 	switch (msec) {
 	default:
-		index = 11;
+		idx = 11;
 		break;
 	case 1000 ... 1999:
-		index = 10;
+		idx = 10;
 		break;
 	case 750 ... 999:
-		index = 9;
+		idx = 9;
 		break;
 	case 500 ... 749:
-		index = 8;
+		idx = 8;
 		break;
 	case 250 ... 499:
-		index = 7;
+		idx = 7;
 		break;
 	case 100 ... 249:
-		index = 6;
+		idx = 6;
 		break;
 	case 50 ... 99:
-		index = 5;
+		idx = 5;
 		break;
 	case 20 ... 49:
-		index = 4;
+		idx = 4;
 		break;
 	case 10 ... 19:
-		index = 3;
+		idx = 3;
 		break;
 	case 4 ... 9:
-		index = 2;
+		idx = 2;
 		break;
 	case 2 ... 3:
-		index = 1;
+		idx = 1;
 	case 0 ... 1:
 		break;
 	}
 
-	assert(index < FIO_IO_U_LAT_M_NR);
-	td->ts.io_u_lat_m[index]++;
+	assert(idx < FIO_IO_U_LAT_M_NR);
+	td->ts.io_u_lat_m[idx]++;
 }
 
 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
@@ -733,7 +887,8 @@ static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
 			opened = 1;
 		}
 
-		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
+		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
+								f->flags);
 		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
 			break;
 
@@ -747,7 +902,7 @@ static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
 	return f;
 }
 
-static struct fio_file *get_next_file(struct thread_data *td)
+static struct fio_file *__get_next_file(struct thread_data *td)
 {
 	struct fio_file *f;
 
@@ -782,6 +937,16 @@ out:
 	return f;
 }
 
+static struct fio_file *get_next_file(struct thread_data *td)
+{
+	struct prof_io_ops *ops = &td->prof_io_ops;
+
+	if (ops->get_next_file)
+		return ops->get_next_file(td);
+
+	return __get_next_file(td);
+}
+
 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
 {
 	struct fio_file *f;
@@ -802,7 +967,8 @@ static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
 		io_u->file = NULL;
 		fio_file_set_done(f);
 		td->nr_done_files++;
-		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files);
+		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
+					td->nr_done_files, td->o.nr_files);
 	} while (1);
 
 	return 0;
@@ -813,6 +979,9 @@ struct io_u *__get_io_u(struct thread_data *td)
 {
 	struct io_u *io_u = NULL;
 
+	td_io_u_lock(td);
+
+again:
 	if (!flist_empty(&td->io_u_requeues))
 		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
 	else if (!queue_full(td)) {
@@ -826,17 +995,73 @@ struct io_u *__get_io_u(struct thread_data *td)
 
 	if (io_u) {
 		assert(io_u->flags & IO_U_F_FREE);
-		io_u->flags &= ~IO_U_F_FREE;
+		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
+		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
 
 		io_u->error = 0;
 		flist_del(&io_u->list);
 		flist_add(&io_u->list, &td->io_u_busylist);
 		td->cur_depth++;
+		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
+	} else if (td->o.verify_async) {
+		/*
+		 * We ran out, wait for async verify threads to finish and
+		 * return one
+		 */
+		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
+		goto again;
 	}
 
+	td_io_u_unlock(td);
 	return io_u;
 }
 
+static int check_get_trim(struct thread_data *td, struct io_u *io_u)
+{
+	if (td->o.trim_backlog && td->trim_entries) {
+		int get_trim = 0;
+
+		if (td->trim_batch) {
+			td->trim_batch--;
+			get_trim = 1;
+		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
+			 td->last_ddir != DDIR_READ) {
+			td->trim_batch = td->o.trim_batch;
+			if (!td->trim_batch)
+				td->trim_batch = td->o.trim_backlog;
+			get_trim = 1;
+		}
+
+		if (get_trim && !get_next_trim(td, io_u))
+			return 1;
+	}
+
+	return 0;
+}
+
+static int check_get_verify(struct thread_data *td, struct io_u *io_u)
+{
+	if (td->o.verify_backlog && td->io_hist_len) {
+		int get_verify = 0;
+
+		if (td->verify_batch) {
+			td->verify_batch--;
+			get_verify = 1;
+		} else if (!(td->io_hist_len % td->o.verify_backlog) &&
+			 td->last_ddir != DDIR_READ) {
+			td->verify_batch = td->o.verify_batch;
+			if (!td->verify_batch)
+				td->verify_batch = td->o.verify_backlog;
+			get_verify = 1;
+		}
+
+		if (get_verify && !get_next_verify(td, io_u))
+			return 1;
+	}
+
+	return 0;
+}
+
 /*
  * Return an io_u to be processed. Gets a buflen and offset, sets direction,
  * etc. The returned io_u is fully ready to be prepped and submitted.
@@ -852,6 +1077,11 @@ struct io_u *get_io_u(struct thread_data *td)
 		return NULL;
 	}
 
+	if (check_get_verify(td, io_u))
+		goto out;
+	if (check_get_trim(td, io_u))
+		goto out;
+
 	/*
 	 * from a requeue, io_u already setup
 	 */
@@ -872,18 +1102,26 @@ struct io_u *get_io_u(struct thread_data *td)
 	f = io_u->file;
 	assert(fio_file_open(f));
 
-	if (io_u->ddir != DDIR_SYNC) {
+	if (ddir_rw(io_u->ddir)) {
 		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
 			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
 			goto err_put;
 		}
 
+		f->last_start = io_u->offset;
 		f->last_pos = io_u->offset + io_u->buflen;
 
 		if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
 			populate_verify_io_u(td, io_u);
 		else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
 			io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
+		else if (io_u->ddir == DDIR_READ) {
+			/*
+			 * Reset the buf_filled parameters so next time if the
+			 * buffer is used for writes it is refilled.
+			 */
+			io_u->buf_filled_len = 0;
+		}
 	}
 
 	/*
@@ -893,6 +1131,7 @@ struct io_u *get_io_u(struct thread_data *td)
 	io_u->xfer_buflen = io_u->buflen;
 
 out:
+	assert(io_u->file);
 	if (!td_io_prep(td, io_u)) {
 		if (!td->o.disable_slat)
 			fio_gettime(&io_u->start_time, NULL);
@@ -906,7 +1145,10 @@ err_put:
 
 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
 {
-	const char *msg[] = { "read", "write", "sync" };
+	const char *msg[] = { "read", "write", "sync", "datasync",
+				"sync_file_range", "wait", "trim" };
+
+
 
 	log_err("fio: io_u error");
 
@@ -930,50 +1172,80 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 	 * initialized, silence that warning.
 	 */
 	unsigned long uninitialized_var(usec);
+	struct fio_file *f;
 
 	dprint_io_u(io_u, "io complete");
 
+	td_io_u_lock(td);
 	assert(io_u->flags & IO_U_F_FLIGHT);
-	io_u->flags &= ~IO_U_F_FLIGHT;
+	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
+	td_io_u_unlock(td);
 
-	if (io_u->ddir == DDIR_SYNC) {
+	if (ddir_sync(io_u->ddir)) {
 		td->last_was_sync = 1;
+		f = io_u->file;
+		if (f) {
+			f->first_write = -1ULL;
+			f->last_write = -1ULL;
+		}
 		return;
 	}
 
 	td->last_was_sync = 0;
+	td->last_ddir = io_u->ddir;
 
-	if (!io_u->error) {
+	if (!io_u->error && ddir_rw(io_u->ddir)) {
 		unsigned int bytes = io_u->buflen - io_u->resid;
 		const enum fio_ddir idx = io_u->ddir;
+		const enum fio_ddir odx = io_u->ddir ^ 1;
 		int ret;
 
 		td->io_blocks[idx]++;
 		td->io_bytes[idx] += bytes;
 		td->this_io_bytes[idx] += bytes;
 
+		if (idx == DDIR_WRITE) {
+			f = io_u->file;
+			if (f) {
+				if (f->first_write == -1ULL ||
+				    io_u->offset < f->first_write)
+					f->first_write = io_u->offset;
+				if (f->last_write == -1ULL ||
+				    ((io_u->offset + bytes) > f->last_write))
+					f->last_write = io_u->offset + bytes;
+			}
+		}
+
 		if (ramp_time_over(td)) {
 			unsigned long uninitialized_var(lusec);
-			unsigned long uninitialized_var(rusec);
 
 			if (!td->o.disable_clat || !td->o.disable_bw)
 				lusec = utime_since(&io_u->issue_time,
 							&icd->time);
-			if (__should_check_rate(td, idx) ||
-			    __should_check_rate(td, idx ^ 1))
-				rusec = utime_since(&io_u->start_time,
-							&icd->time);
+			if (!td->o.disable_lat) {
+				unsigned long tusec;
 
+				tusec = utime_since(&io_u->start_time,
+							&icd->time);
+				add_lat_sample(td, idx, tusec, bytes);
+			}
 			if (!td->o.disable_clat) {
-				add_clat_sample(td, idx, usec, bytes);
+				add_clat_sample(td, idx, lusec, bytes);
 				io_u_mark_latency(td, lusec);
 			}
 			if (!td->o.disable_bw)
 				add_bw_sample(td, idx, bytes, &icd->time);
-			if (__should_check_rate(td, idx))
-				td->rate_pending_usleep[idx] += (long) td->rate_usec_cycle[idx] - rusec;
+			if (__should_check_rate(td, idx)) {
+				td->rate_pending_usleep[idx] =
+					((td->this_io_bytes[idx] *
+					  td->rate_nsec_cycle[idx]) / 1000 -
+					 utime_since_now(&td->start));
+			}
 			if (__should_check_rate(td, idx ^ 1))
-				td->rate_pending_usleep[idx ^ 1] -= lusec;
+				td->rate_pending_usleep[odx] =
+					((td->this_io_bytes[odx] *
+					  td->rate_nsec_cycle[odx]) / 1000 -
+					 utime_since_now(&td->start));
 		}
 
 		if (td_write(td) && idx == DDIR_WRITE &&
@@ -988,10 +1260,21 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 			if (ret && !icd->error)
 				icd->error = ret;
 		}
-	} else {
+	} else if (io_u->error) {
 		icd->error = io_u->error;
 		io_u_log_error(td, io_u);
 	}
+	if (td->o.continue_on_error && icd->error &&
+	    td_non_fatal_error(icd->error)) {
+		/*
+		 * If there is a non_fatal error, then add to the error count
+		 * and clear all the errors.
+		 */
+		update_error_count(td, icd->error);
+		td_clear_error(td);
+		icd->error = 0;
+		io_u->error = 0;
+	}
 }
 
 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
@@ -1016,7 +1299,9 @@ static void ios_completed(struct thread_data *td,
 		io_u = td->io_ops->event(td, i);
 
 		io_completed(td, io_u, icd);
-		put_io_u(td, io_u);
+
+		if (!(io_u->flags & IO_U_F_FREE_DEF))
+			put_io_u(td, io_u);
 	}
 }
 
@@ -1030,7 +1315,9 @@ int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
 
 	init_icd(td, &icd, 1);
 	io_completed(td, io_u, &icd);
-	put_io_u(td, io_u);
+
+	if (!(io_u->flags & IO_U_F_FREE_DEF))
+		put_io_u(td, io_u);
 
 	if (icd.error) {
 		td_verror(td, icd.error, "io_u_sync_complete");
@@ -1102,13 +1389,10 @@ void io_u_queued(struct thread_data *td, struct io_u *io_u)
 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
 		      unsigned int max_bs)
 {
-	long *ptr = io_u->buf;
+	io_u->buf_filled_len = 0;
 
-	if (!td->o.zero_buffers) {
-		while ((void *) ptr - io_u->buf < max_bs) {
-			*ptr = rand() * GOLDEN_RATIO_PRIME;
-			ptr++;
-		}
-	} else
-		memset(ptr, 0, max_bs);
+	if (!td->o.zero_buffers)
+		fill_random_buf(io_u->buf, max_bs);
+	else
+		memset(io_u->buf, 0, max_bs);
 }